Starling Detector Notes
Vlad Obo
10/6/2018
source("birdsLib.R")##
## Attaching package: 'monitoR'## The following object is masked from 'package:tuneR':
##
## readMP3##
## Attaching package: 'dplyr'## The following objects are masked from 'package:stats':
##
## filter, lag## The following objects are masked from 'package:base':
##
## intersect, setdiff, setequal, unionAbout this document
Problem Description
We have massive sound records and need to find if there are starling calls
There are a data set of starling calls 1000+ records
We think the data in hands reasonably cover the real conditions
we develop specific application for starlings
we consider mel frequency cepstrum coeffisients as a feature and HTK and KERAS as ML tools
Parts
Choosing MFCC parameters
After playing the data the following numbers were chosen as MFCC’s parameters.
Happy to explaine the choice.
hopt## [1] 0.025wint ## [1] 0.075getMel## function (x)
## {
## require(tuneR)
## tmp <- melfcc(x, minfreq = 3000, maxfreq = 10000, fbtype = "mel",
## numcep = 12, hopt = hopt, wint = wint)
## return(tmp)
## }Bird call detector
Simple detector to cut birds calls from lengthy sound records would be good to have
I called it “MFCC power”
The function is as below
getMCPow## function (x, excl = 1)
## {
## colSums(t(abs(x[, -(excl)])))
## }- How it works. Using 15 seconds from joined starling decreasin calls as an example
sigDir="../"
fn="DecreasingCallLargeTmp.wav"
w<-readWave(paste0(sigDir, fn), from = 0, to = 15, units = "sec")
plot(w)
viewSpec(w)
ss<-getMel(w)
tt<-(1:nrow(ss)) * hopt
matplot(tt, ss, type="l", main="MFCCs, all")
pov=getMCPow(ss)
plot(pov, type="b", main="MFFC Power")
showMCPow(pov)
title(main="Median filter")
win.range<-c(0.4, 0.8)I use this file DecreasingCallLargeTmp.wav to produce the estimate for target window length
The target window size set to 0.4, 0.8 sec
wins<-readWave(paste0(sigDir, fn)) %>%
getMel %>%
getMCPow %>%
getAWins
ll<-with(wins, end-start)
plot(sort(ll))
abline(h=win.range, lty=2, col="red")
Real audio
Now worrking on the file MU000_20091021_172000.wav (the first helf hour only here to avoid memory problems)
From 0 to 30, in minutes
Action windows in real audio file
We seek time windows with enough action in them
It means when the median filter output is greater than the threshold
Threshold used by default (see function “getAWins”) is the top quartile of “pov”
## be careful -- remove the tmp file when changing the input wav
tmp.fn="action.wins.RData"
#file.remove(tmp.fn)
if (file.exists(tmp.fn)) load(tmp.fn) else {
action.wins <- (ppov<- paste0(sigDir, lfn) %>%
readWave(from=0, to = 30, units="min") %>%
mono %>%
getMel() %>%
getMCPow()
) %>%
getAWins()
save(ppov, action.wins, file=tmp.fn) }
ll<-with(action.wins, end-start)
plot(sort(ll), log="y")
abline(h=win.range, col="red", lty=2)
ii<-which(inrange(ll, win.range))- Number of action windows similar to starling decr.call window size
table(inrange(ll, win.range))##
## FALSE TRUE
## 740 292- Total sum of lengths for the windows of these sizes in sec (this is from the record of 1h)
sum(ll[ii])## [1] 161.875- Let’s generate wav including only fragments determined by the action windows. Accomponying selection table is written too. The can be viewed by Raven
ofn="wins.wav"
winsWav(infile=ifn, outfile=ofn, ww=action.wins[ii,])## Warning in .local(left, ...): 'samp.rate' not specified, assuming 44100Hz## Warning in .local(left, ...): 'bit' not specified, assuming 16bit- Selection table to be used with original file to show these selected windows
writeSelectionTable(wins=action.wins, ofn="actionWins.txt")Test file
sigDir<-"/Volumes/VladBackup/Birds/StarlingCalls/"
lfn<-"ma5ho-ehcz5.L.wav"
wL<-mono(readWave(paste0(sigDir, lfn)))
plot(wL, main=lfn)
viewSpec(wL, main=lfn)
ss=getMel(wL)
pov=getMCPow(ss)
showMCPow(pov, tit=lfn)
xll=c(0,300)
op<-par(mfcol=c(3,1), mar=c(0,1,0,1))
viewSpec(wL, start.time = xll[1], page.length = xll[2])
showMCPow(pov, tit=lfn, xl = xll)
plot(wL, xlim=xll)